A compound framework incorporating improved outlier detection and correction, VMD, weight-based stacked generalization with enhanced DESMA for multi-step short-term wind speed forecasting

Abstract

Precise wind speed forecasting contributes to wind power consumption and power grid schedule as well as promotes the implementation of global carbon neutrality policy. However, in existing research, the negative impact of outliers on forecasting models is ignored and the inherent shortcomings of the single predictors have not been taken seriously. Moreover, the intrinsic parameters of predictors are set by manual and empirical methods in some research, leading to difficulties in achieving optimal forecasting performance. To solve the shortcomings of existing research, a multi-step short-term wind speed forecasting framework is proposed by incorporating boxplot-medcouple (MC), variational mode decomposition (VMD), phase space reconstruction (PSR), weight-based stacked generalization with enhanced differential evolution slime mold algorithm (DESMA). Firstly, boxplot-MC is employed to achieve outlier detection and correction for preprocessing original wind speed data by analyzing values and trends. Then, the modified data is further adaptively decomposed into multiple subsequences by VMD, after which each subsequence is constructed into feature matrices through PSR. Subsequently, weight-based multi-model fusion strategy in layer-1 of stacked generalization is proposed to integrate the predicting values acquired by three primary learners, of which the weight coefficients are calculated with the error between actual values and predicting values. After that, kernel extreme learning machine (KELM) in layer-2 of stacked generalization is applied to predict the fusion result to obtain forecasting value corresponding to each subsequence. Meanwhile, an enhanced DESMA based on slime mold algorithm (SMA) and differential evolution (DE) is proposed to calibrate the parameters of KELM. Eventually, the final wind speed forecasting results are attained by summing the prediction values of all subsequences. Furthermore, comparative experiments from different aspects are undertaken on real datasets to ascertain the availability of the proposed framework. The experimental results are clarified as follows: (1) outlier detection and correction employing boxplot-MC is dedicated to analyzing values and trends effectively, with which the negative impact of outliers can be weakened while retaining valid data significantly; (2) VMD can prominently reduce the non-smoothness and volatility of wind speed data; (3) weight-based stacked generalization is conducive to exploiting the advantages of individual primary learners, contributing to compensating for instability; (4) DESMA enhances prediction accuracy by optimizing the parameters of KELM. Additionally, the code has been made available at https://github.com/fyc233/a-multi-step-short-term-wind-speed-forecasting-framework.git.